Yingfei Xiong

Trae Research Team, Pengfei Gao, Zhao Tian et al. · pku

Software issue resolution is a critical challenge in software engineering and has garnered increasing attention in recent years. With the rapid advancement of large language models (LLMs), substantial progress has been made in addressing real-world software engineering tasks. Recent studies have introduced ensemble reasoning techniques to enhance the performance of LLM-based issue resolution. However, existing prompting-based methods still face limitations in effectively exploring large ensemble spaces and lack the capacity for repository-level understanding, both of which constrain their overall effectiveness. In this paper, we propose Trae Agent, the first agent-based ensemble reasoning approach for repository-level issue resolution. Trae Agent formulates our goal as an optimal solution search problem and addresses two key challenges, i.e., large ensemble spaces and repository-level understanding, through modular agents for generation, pruning, and selection. We conduct extensive experiments using three leading LLMs on the widely-adopted SWE-bench benchmark, comparing Trae Agent against four state-of-the-art ensemble reasoning techniques. Experimental results demonstrate that Trae Agent consistently achieves superior performance, with an average improvement of 10.22% over all baselines in terms of Pass@1. Trae Agent has achieved first place on the SWE-bench Verified leaderboard, with a notable Pass@1 score of 75.20%. We are pleased to release Trae Agent as an open-source project to support the research community, with all resources available at https://github.com/bytedance/trae-agent.

Daya Guo, Qihao Zhu, Dejian Yang et al.

The rapid development of large language models has revolutionized code intelligence in software development. However, the predominance of closed-source models has restricted extensive research and development. To address this, we introduce the DeepSeek-Coder series, a range of open-source code models with sizes from 1.3B to 33B, trained from scratch on 2 trillion tokens. These models are pre-trained on a high-quality project-level code corpus and employ a fill-in-the-blank task with a 16K window to enhance code generation and infilling. Our extensive evaluations demonstrate that DeepSeek-Coder not only achieves state-of-the-art performance among open-source code models across multiple benchmarks but also surpasses existing closed-source models like Codex and GPT-3.5. Furthermore, DeepSeek-Coder models are under a permissive license that allows for both research and unrestricted commercial use.

Qihao Zhu, Zeyu Sun, Xiran Liang et al.

Code retrieval helps developers reuse the code snippet in the open-source projects. Given a natural language description, code retrieval aims to search for the most relevant code among a set of code. Existing state-of-the-art approaches apply neural networks to code retrieval. However, these approaches still fail to capture an important feature: overlaps. The overlaps between different names used by different people indicate that two different names may be potentially related (e.g., "message" and "msg"), and the overlaps between identifiers in code and words in natural language descriptions indicate that the code snippet and the description may potentially be related. To address these problems, we propose a novel neural architecture named OCoR, where we introduce two specifically-designed components to capture overlaps: the first embeds identifiers by character to capture the overlaps between identifiers, and the second introduces a novel overlap matrix to represent the degrees of overlaps between each natural language word and each identifier. The evaluation was conducted on two established datasets. The experimental results show that OCoR significantly outperforms the existing state-of-the-art approaches and achieves 13.1% to 22.3% improvements. Moreover, we also conducted several in-depth experiments to help understand the performance of different components in OCoR.

Jingjing Liang, Ruyi Ji, Jiajun Jiang et al.

It is widely recognized that program repair tools need to have a high precision to be useful, i.e., the generated patches need to have a high probability to be correct. However, it is fundamentally difficult to ensure the correctness of the patches, and many tools compromise other aspects of repair performance such as recall for an acceptable precision. In this paper we ask a question: can a repair tool with a low precision be still useful? To explore this question, we propose an interactive filtering approach to patch review, which filters out incorrect patches by asking questions to the developers. Our intuition is that incorrect patches can still help understand the bug. With proper tool support, the benefit outweighs the cost even if there are many incorrect patches. We implemented the approach as an Eclipse plugin tool, InPaFer, and evaluated it with a simulated experiment and a user study with 30 developers. The results show that our approach improve the repair performance of developers, with 62.5% more successfully repaired bugs and 25.3% less debugging time in average. In particular, even if the generated patches are all incorrect, the performance of the developers would not be significantly reduced, and could be improved when some patches provide useful information for repairing, such as the faulty location and a partial fix.

Zeyu Sun, Qihao Zhu, Yingfei Xiong et al.

A code generation system generates programming language code based on an input natural language description. State-of-the-art approaches rely on neural networks for code generation. However, these code generators suffer from two problems. One is the long dependency problem, where a code element often depends on another far-away code element. A variable reference, for example, depends on its definition, which may appear quite a few lines before. The other problem is structure modeling, as programs contain rich structural information. In this paper, we propose a novel tree-based neural architecture, TreeGen, for code generation. TreeGen uses the attention mechanism of Transformers to alleviate the long-dependency problem, and introduces a novel AST reader (encoder) to incorporate grammar rules and AST structures into the network. We evaluated TreeGen on a Python benchmark, HearthStone, and two semantic parsing benchmarks, ATIS and GEO. TreeGen outperformed the previous state-of-the-art approach by 4.5 percentage points on HearthStone, and achieved the best accuracy among neural network-based approaches on ATIS (89.1%) and GEO (89.6%). We also conducted an ablation test to better understand each component of our model.

Zeyu Sun, Qihao Zhu, Lili Mou et al.

Code generation maps a program description to executable source code in a programming language. Existing approaches mainly rely on a recurrent neural network (RNN) as the decoder. However, we find that a program contains significantly more tokens than a natural language sentence, and thus it may be inappropriate for RNN to capture such a long sequence. In this paper, we propose a grammar-based structural convolutional neural network (CNN) for code generation. Our model generates a program by predicting the grammar rules of the programming language; we design several CNN modules, including the tree-based convolution and pre-order convolution, whose information is further aggregated by dedicated attentive pooling layers. Experimental results on the HearthStone benchmark dataset show that our CNN code generator significantly outperforms the previous state-of-the-art method by 5 percentage points; additional experiments on several semantic parsing tasks demonstrate the robustness of our model. We also conduct in-depth ablation test to better understand each component of our model.

Ruru Yue, Zhe Gao, Na Meng et al.

When many clones are detected in software programs, not all clones are equally important to developers. To help developers refactor code and improve software quality, various tools were built to recommend clone-removal refactorings based on the past and the present information, such as the cohesion degree of individual clones or the co-evolution relations of clone peers. The existence of these tools inspired us to build an approach that considers as many factors as possible to more accurately recommend clones. This paper introduces CREC, a learning-based approach that recommends clones by extracting features from the current status and past history of software projects. Given a set of software repositories, CREC first automatically extracts the clone groups historically refactored (R-clones) and those not refactored (NR-clones) to construct the training set. CREC extracts 34 features to characterize the content and evolution behaviors of individual clones, as well as the spatial, syntactical, and co-change relations of clone peers. With these features, CREC trains a classifier that recommends clones for refactoring. We designed the largest feature set thus far for clone recommendation, and performed an evaluation on six large projects. The results show that our approach suggested refactorings with 83% and 76% F-scores in the within-project and cross-project settings. CREC significantly outperforms a state-of-the-art similar approach on our data set, with the latter one achieving 70% and 50% F-scores. We also compared the effectiveness of different factors and different learning algorithms.

Daming Zou, Jingjing Liang, Yingfei Xiong et al.

The performance of fault localization techniques is critical to their adoption in practice. This paper reports on an empirical study of a wide range of fault localization techniques on real-world faults. Different from previous studies, this paper (1) considers a wide range of techniques from different families, (2) combines different techniques, and (3) considers the execution time of different techniques. Our results reveal that a combined technique significantly outperforms any individual technique (200% increase in faults localized in Top 1), suggesting that combination may be a desirable way to apply fault localization techniques and that future techniques should also be evaluated in the combined setting. Our implementation is publicly available for evaluating and combining fault localization techniques.

Yingfei Xiong, Bo Wang, Guirong Fu et al.

In many scenarios we need to find the most likely program under a local context, where the local context can be an incomplete program, a partial specification, natural language description, etc. We call such problem program estimation. In this paper we propose an abstract framework, learning to synthesis, or L2S in short, to address this problem. L2S combines four tools to achieve this: syntax is used to define the search space and search steps, constraints are used to prune off invalid candidates at each search step, machine-learned models are used to estimate conditional probabilities for the candidates at each search step, and search algorithms are used to find the best possible solution. The main goal of L2S is to lay out the design space to motivate the research on program estimation. We have performed a preliminary evaluation by instantiating this framework for synthesizing conditions of an automated program repair (APR) system. The training data are from the project itself and related JDK packages. Compared to ACS, a state-of-the-art condition synthesis system for program repair, our approach could deal with a larger search space such that we fixed 4 additional bugs outside the search space of ACS, and relies only on the source code of the current projects.

Yingfei Xiong, Xinyuan Liu, Muhan Zeng et al.

Test-based automatic program repair has attracted a lot of attention in recent years. However, the test suites in practice are often too weak to guarantee correctness and existing approaches often generate a large number of incorrect patches. To reduce the number of incorrect patches generated, we propose a novel approach that heuristically determines the correctness of the generated patches. The core idea is to exploit the behavior similarity of test case executions. The passing tests on original and patched programs are likely to behave similarly while the failing tests on original and patched programs are likely to behave differently. Also, if two tests exhibit similar runtime behavior, the two tests are likely to have the same test results. Based on these observations, we generate new test inputs to enhance the test suites and use their behavior similarity to determine patch correctness. Our approach is evaluated on a dataset consisting of 139 patches generated from existing program repair systems including jGenProg, Nopol, jKali, ACS and HDRepair. Our approach successfully prevented 56.3\% of the incorrect patches to be generated, without blocking any correct patches.

Jiajun Jiang, Yingfei Xiong

Automated program repair techniques, which target to generating correct patches for real world defects automatically, have gained a lot of attention in the last decade. Many different techniques and tools have been proposed and developed. However, even the most sophisticated program repair techniques can only repair a small portion of defects while producing a lot of incorrect patches. A possible reason for this low performance is that the test suites of real world programs are usually too weak to guarantee the behavior of the program. To understand to what extent defects can be fixed with weak test suites, we analyzed 50 real world defects from Defects4J, in which we found that up to 84% of them could be correctly fixed. This result suggests that there is plenty of space for current automated program repair techniques to improve. Furthermore, we summarized seven fault localization strategies and seven patch generation strategies that were useful in localizing and fixing these defects, and compared those strategies with current repair techniques. The results indicate potential directions to improve automatic program repair in the future research.

Bo Wang, Yingfei Xiong, Yangqingwei Shi et al.

Mutation analysis has many applications, such as asserting the quality of test suites and localizing faults. One important bottleneck of mutation analysis is scalability. The latest work explores the possibility of reducing the redundant execution via split-stream execution. However, split-stream execution is only able to remove redundant execution before the first mutated statement. In this paper we try to also reduce some of the redundant execution after the execution of the first mutated statement. We observe that, although many mutated statements are not equivalent, the execution result of those mutated statements may still be equivalent to the result of the original statement. In other words, the statements are equivalent modulo the current state. In this paper we propose a fast mutation analysis approach, AccMut. AccMut automatically detects the equivalence modulo states among a statement and its mutations, then groups the statements into equivalence classes modulo states, and uses only one process to represent each class. In this way, we can significantly reduce the number of split processes. Our experiments show that our approach can further accelerate mutation analysis on top of split-stream execution with a speedup of 2.56x on average.

Yingfei Xiong, Jie Wang, Runfa Yan et al.

Due to the difficulty of repairing defect, many research efforts have been devoted into automatic defect repair. Given a buggy program that fails some test cases, a typical automatic repair technique tries to modify the program to make all tests pass. However, since the test suites in real world projects are usually insufficient, aiming at passing the test suites often leads to incorrect patches. In this paper we aim to produce precise patches, that is, any patch we produce has a relatively high probability to be correct. More concretely, we focus on condition synthesis, which was shown to be able to repair more than half of the defects in existing approaches. Our key insight is threefold. First, it is important to know what variables in a local context should be used in an "if" condition, and we propose a sorting method based on the dependency relations between variables. Second, we observe that the API document can be used to guide the repair process, and propose document analysis technique to further filter the variables. Third, it is important to know what predicates should be performed on the set of variables, and we propose to mine a set of frequently used predicates in similar contexts from existing projects. We develop a novel program repair system, ACS, that could generate precise conditions at faulty locations. Furthermore, given the generated conditions are very precise, we can perform a repair operation that is previously deemed to be too overfitting: directly returning the test oracle to repair the defect. Using our approach, we successfully repaired 17 defects on four projects of Defects4J, which is the largest number of fully automatically repaired defects reported on the dataset so far. More importantly, the precision of our approach in the evaluation is 73.9%, which is significantly higher than previous approaches, which are usually less than 40%.